Machine Learning Compute Rig

Dash Lambda · January 2, 2016

With my desktop nearing the point where I no longer have anything I want to upgrade (Only the storage left...), I'm planning another PC.

I had a similar topic a while ago, but a lot about both my understanding of the hardware and the hardware itself has changed. I'm looking for advice primarily about the CPU and GPU.

The purpose of this rig is in the title, it will be used for developing neural networks.

As for details on my situation, I am a student and will be using this rig while learning- I am not a professional doing proper research. That also means I don't have the money for Tesla K80's or 2699-v3's.

Anyway, on to the parts...

CPU: As most of the work will be done on the graphics cards, I don't imagine I'll need too beefy a CPU. I do want a Xeon, preferably 6-8 cores. My issue here is that I don't know the first thing about Xeons. This is my first dive into workstation/server hardware, so I don't know what to choose.

GPU: I'll be using CUDA, so it logically follows that I'll want to go with Nvidia cards. I have two concerns:

-Should I use Kepler, Maxwell, or wait for Pascal? I've heard that Kepler is actually faster for deep learning than Maxwell, but then of course Pascal promises a "10x speedup in mixed-precision workloads".

-Should I use consumer or professional cards? Would a Titan work just as well as a Quadro? It's my understanding that the difference is stability, power consumption, and heat output- None of which, to a point, are particularly important to me (With the possible exception of stability. How big is the difference?).

Also, I plan to run this system in a Caselabs Mercury S8 test-bench. Is an open case a bad idea?

EDIT: I almost forgot to mention, I also know nothing about workstation/server motherboards. Any suggestions?

Streetguru · January 2, 2016

With my desktop nearing the point where I no longer have anything I want to upgrade (Only the storage left...), I'm planning another PC.

budget/country

you might as well wait for pascal at this point, although by that time AMD's CUDA translator thing might be out and in use possibly making their cards the better choice.

Otherwise you'll probably just want to go X99 for the CPU